Aggregates containing elements with constexpr constructors are not statically initalized

[BillyONeal]

Bugzilla Link	45367
Version	10.0
OS	Windows NT
Attachments	repro.cpp
CC	@efriedma-quic,@saxbophone,@zygoloid,@rnk,@yurybura

Extended Description

Previously reported as microsoft/STL#661

In the following example, should_be_statically_initalized should be eligible for static initialization, then the dynamic initializer for foo that calls bar() should run, resulting in the asserts in main(). However, clang only seems to give these static initializers if the varaible is {}'d.

C:\Users\billy\Desktop>type repro.cpp

#include <assert.h>

struct FakeAtomic {
    int example;

    constexpr FakeAtomic() noexcept : example{} {}
    int operator++() { return ++example; }
    int operator++(int) { return example++; }
    int load() const { return example; }
};

struct Meow {
    FakeAtomic data[2];
    char padding[64];
};

int bar();

int foo = bar();
#ifdef WORKAROUND
Meow should_be_statically_initalized[256]{};
#else
Meow should_be_statically_initalized[256];
#endif

int bar() { return should_be_statically_initalized[0].data[0]++; }

int main() {
    assert(foo == 0);
    assert(should_be_statically_initalized[0].data[0].load() == 1);
}

C:\Users\billy\Desktop>clang-cl /EHsc /W4 /WX .\repro.cpp

C:\Users\billy\Desktop>.\repro.exe
Assertion failed: should_be_statically_initalized[0].data[0].load() == 1, file .\repro.cpp, line 30

C:\Users\billy\Desktop>clang-cl /DWORKAROUND /EHsc /W4 /WX .\repro.cpp

C:\Users\billy\Desktop>.\repro.exe

C:\Users\billy\Desktop>

[BillyONeal]

Hmmm rereading [basic.start.static]/2 it says that either constant initialization is performed which would require an initializer, or zero initialization is performed, but we don't seem to apply that recursively. So clang may be correct here, but if it is, I'd argue that this would be a CWG issue of some sort.

[efriedma-quic]

Default initialization on a "Meow" doesn't produce a constant: the "padding" is uninitialized, and a value with uninitialized fields isn't a constant. List initialization on a "Meow" does produce a constant: "padding" is initialized to zero.

Given that, I think clang is correctly interpreting what it means for a variable to be "constant-initialized". But it could be more aggressive about using static initialization anyway.

[zygoloid]

This example was broken by the resolution of CWG2026 (http://wg21.link/cwg2026), prior to which we would perform zero-initialization before considering the initializer for this global, as we would when performing dynamic initialization. (Clang 10 doesn't implement CWG2026, which is why it accepts. Clang < 10 reject because they don't implement the C++20 change P1331R2, which is what allows Meow to have a constexpr default constructor despite not initializing all its members.)

[frederick-vs-ja]

This seems largely related to CWG2536 to me and thus perhaps should be tagged waiting-for-wg21.